N-Dimensional X-Ray Image for Lungs Abnormalities Detection Using Deep Learning Technique

Abstract

This paper presents a methodology to generate an N-dimensional, stacked image dataset. The chest X-Ray images dataset, acquired from the NIH database is used to develop an N-dimensional images dataset. The NIH published a list of chest X-ray images to aid the scientific community in the research work. The dataset consists of a large number of X-ray images of multiple chest diseases. In this work, we selected 3,500 images, distributed equally to the five chest findings. The dataset generation process involves suppressing undesired distortions and enhancing desired features on the radiological images. The feature enhancement is achieved by applying multiple filters on the classical digital X-ray images. The multi-filtering technique aims to enhance and elaborate abnormalities in the image, letting the classification models detect and extract slight variations in the image features. The presented work aims to get an improved classification result on chest diseases to help decision-making easy and less erroneous. The preprocessed dataset is then fed to the deep convolutional neural network (CNN) models like VGG16, ResNet, and Inception. The models are custom-tailored to accept N-dimensional stacked images and transfer learning is also applied to the models thus eliminating the need for retraining the models. A lightweight deep CNN model is also designed to feature considerably fewer layers and weights. The model is quickly trained on underpowered devices. The two model sets are then evaluated to detect and classify the chest disease in the formulated images. The evaluations are applied to chest X-ray images of multiple classes. The experimental results that the models applied to the proposed multichannel image dataset showed 95% higher classification accuracy than the experiment results from the original X-ray image dataset.